Commit Graph
168 Commits
Author SHA1 Message Date
maximeandClaude Opus 5.5 1ced7667b1 Detector-learner: shorter benchmark per station, work folders, grouped comparisons, control journal at base
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:35:56 +01:00
maximeandClaude Opus 5.5 c14781698a Detector-learner: process fixes after the Obstruction session (repumped data, sorts in capteurs.py, critic folder)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 12:12:20 +01:00
maximeandClaude Opus 5.5 d1a372d3ea Detector-learner: detectors do their best, a coherent alternative reading is no regression; process fixes (D-135)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 10:54:04 +01:00
maximeandClaude Opus 5.5 eb61d12b82 Detectors read the camera pose and the station metrics when they declare them (D-134)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 10:43:35 +01:00
maximeandClaude Opus 5.5 be147e92df Detector-learner per family, on all stations, launched by Claude (D-132)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 09:13:48 +01:00
maximeandClaude Opus 5.5 c6e9df0e46 Stable ids for auto labels, a summary for the label map, reading tools improved (D-131)
- Every auto label carries an id (<STATION>-<n>, labeling/ids.py), kept on edit, given to
  new labels by labeling.py and by the reviews; controle.py requires it, unique. The
  cross-review matches labels by id, and an applied review is locked (applique.md).
- The label map starts with a summary (labels-map/sommaire.py): readers take the summary and
  sections 2-4 in full, then only the cards of the phenomena they meet.
- Tool-builder session on the reviewers' difficulties: jours.py --lissage/--fraction,
  ensemble.py --suivre and rejected measures left out of --rigide, --residus non, zero noise
  after a lost reference, new agent-tools/stations.py (several stations on one timeline),
  long outputs convention in lire-les-donnees.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 08:49:54 +01:00
maximeandClaude Opus 5.5 d55de98746 Labeling processes: lessons from the auto-label reviews (D-130)
- The labeler writes to its own folder (etiquettes.jsonl); labeling/auto/ only holds
  reviewed labels, merged by /review-auto-labels (new apply script: dated, sorted, checked).
- labeling/controle.py also checks a proposed file; the labeler and reviewers run it before
  handing back. Reviewers keep their measurements in revue.md, never in the labels.
- Reviewer instructions explicitly add their scripts, .venv/bin/python and for loops to the
  lire-les-donnees commands (the contradiction raised by 12 reviews).
- Expert validation optional in /review-auto-labels; an expert hint given to the labeler is
  written in the labels it touches. Cross-review ends with the releveur.
- docs/ameliorations-processus.md: 17 entries from this review, by the releveur.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 08:31:53 +01:00
maximeandClaude Opus 5.5 35b5ed4d74 Auto labels reviewed per station and cross-reviewed; all labeling knowledge in labeling/ (D-130)
- Every station's auto labels reviewed by an analyst (/review-auto-labels), then a
  cross-review per family across all stations (new skill /cross-review-labels),
  a finishing pass, and descriptions rewritten in words (no numeric values).
- The three never-labeled ranges (WILDSTRUBEL 07/09-01/10, NORDEND 16/08-01/10)
  labeled and reviewed: 991 auto labels on 11 stations, labeling/controle.py at 0 errors.
- Labeling knowledge gathered in labeling/: regles-d-etiquetage.md (new), familles.md
  (settled boundaries), connaissances/ (notes moved from docs/, links updated),
  the label map rebuilt from reviewed labels, events and cases; events tied to their
  stations (PS_DES_COUDRES is JUNGFRAU). The labeler and reviewers read only labeling/.
- labels-map reads auto labels and events; decision D-130; living plan points 73-75.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 08:21:57 +01:00
maximeandClaude Opus 5.5 a2b1f2ed14 Image quality metrics in labeling: qualite.py, LENZSPITZE October labels reviewed with them (D-129)
Definitions of peak, flux, sharpness, fwhm, dominance and snr in docs/domaine;
agent-tools/qualite.py (ratio to a calm reference, per label, outside labels,
figures), inventaire/graphique/resume aware of the six columns, lire-les-donnees
and the labeler's instructions updated. LENZSPITZE 16.10-01.11 relabeled: 67
labels, 4 found by the metrics.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 13:34:04 +01:00
maximeandClaude Opus 5.5 295c856ec2 Coherence of BREITHORN, MICHIGAN and NEBRASKA October labels (same building, per the expert)
One analyst checks neighbouring stations together (.claude/skills/review-auto-labels/coherence.md). The month's drift, shared by the three cameras (same step on 08/10, same recoil 13-16/10), is Displacement everywhere, cause not established (the building or its temperature); the 09/10 reference losses and the 15-16/10 gap are tied across stations, a common cause to look for in the building; the 26/10 cold morning added on BREITHORN and NEBRASKA; MICHIGAN 27/10 as Displacement. Proposed per-station files versioned.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 09:28:19 +01:00
maximeandClaude Opus 5.5 208db519ab Labeler: lessons from the three October reviews (BREITHORN, MICHIGAN, NEBRASKA)
Bounds checked target by target on both columns; a slowing drift stays one event; a phenomenon found once is searched over the whole batch; "the camera, not the targets" checked with ensemble.py --rigide; housing temperature, "d'après l'expert" in words. familles.md: a small error_value rise at the housing temperature trough, nothing on the images, is Thermal displacement, not Visibility. Map lessons left to the next /labels-map.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 09:06:44 +01:00
maximeandClaude Opus 5.5 cef7754888 Reviewed labels: "revue" is the review date, as for manual labels, and is not reviewed again
All 26 BREITHORN October labels carry "revue": "2026-11-04"; the expert's thermal answer is said in the description. /review-auto-labels skips labels that already carry "revue" (count, analyst review, replacement); /labels-map reads all reviewed auto labels (D-128).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 08:46:06 +01:00
maximeandClaude Opus 5.5 83479499a9 Review of the labeler's labels: /review-auto-labels (D-128); BREITHORN October reviewed
The analyst reviews labeling/auto/ labels on a range chosen by the user (bounds, one event, family, description, missed ones) and proposes corrected labels; the expert answers its questions, then the range is replaced. First run on BREITHORN October: 10 corrected, 25 and 26 merged, one missed night added; the expert confirms the camera's cold-night deviations are thermal (housing temperature, not air). Labels the expert settled carry "revue", read by /labels-map.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 08:40:27 +01:00
maximeandClaude Opus 5.5 fab7ddd9fd Presentations: the lost reference is the message (MICHIGAN, NEBRASKA); BREITHORN parallax in corr_y
MICHIGAN and NEBRASKA: the reference network was complete at installation and
stable; since a reference disappeared, the network is altered and the curves
are more agitated (MICHIGAN std 0.014 -> 0.10 px; NEBRASKA 0.016-0.018 ->
0.031 px from 17.10). Later agitations shown as its consequence; MICHIGAN's
too-close targets moved to the end. BREITHORN slide 4 in corr_y, COLI24's
descent stated as unexplained.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 17:49:34 +01:00
maximeandClaude Opus 5.5 8b53df6c50 Presentations: corr_y is in pixels, not mm
Texts and figure axes of the three October decks, figure-y.py label, and a
reminder in the presentation skill.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 17:39:03 +01:00
maximeandClaude Opus 5.5 85e590762b Presentations for the client: Sandouville, no data-gap slide, settlements in corr_y (D-127)
The three October decks (BREITHORN, MICHIGAN, NEBRASKA) no longer criticise
the system: facts and their external cause only. Figures redrawn in corr_y
(figure-y.py), BREITHORN parallax kept. Sources and PDFs named after the
station; the charter's draft stamp reads "Sandouville".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 17:36:25 +01:00
maximeandClaude Opus 5.5 efbc5f15d7 Reading tools: what the etiqueteur had to script itself (metrics, daily bounds, local level, zone, references)
- metriques.py: new, reads metrics/ (temperatures, voltage) against targets and camera pose
- jours.py --bornes: per-day first/last excursion, extreme, from the previous night
- plages.py --glissant, --instants: deviations from a local level, per-instant view
- ensemble.py --table, resume.py --marque/--instants: shorter outputs, repeated headers
- inventaire.py: image gaps as runs, references on reproj_err when inlier is empty
- vignettes.py --zone: same scene area across hours
- donnees.py: references recognised when pattern_type is empty (benchmark unchanged)
- lire-les-donnees: new options documented, gif.py row added

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 16:56:59 +01:00
maximeandClaude Opus 5.5 6fbd7510b7 Etiqueteur: October labels for 8 stations; its folder moves to labeling/auto/etiqueteur/, analyses and scripts versioned (D-126)
Labels for CLARIDEN, FLORIDA, JUNGFRAU (gaps 06-20.10, 26.10-01.11), LENZSPITZE,
NIEDERHORN, NORDEND, SNCF01, WILDSTRUBEL. Figures and tool outputs stay out of git.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 16:33:41 +01:00
maximeandClaude Opus 5.5 2bba388290 Detectors receive the station temperature; ISO dates accepted (D-124)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:56:30 +01:00
maximeandClaude Opus 5.5 daa1fd989d Detector-learner: label bounds excluded, false positives as a regression, control watches the session station (D-123)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:45:13 +01:00
maximeandClaude Opus 5.5 e39531fd33 Detector-learner: user decisions apart in the bilan, process agents without events-learner (D-122)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:21:29 +01:00
maximeandClaude Opus 5.5 6d78bac736 Detector-learner: one commit per suggestion, retirer verdict, measured regression (D-121)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:19:23 +01:00
maximeandClaude Opus 5.5 e8e6cdd96d JUNGFRAU labels get their family; lone jumps scored as Displacement (D-120)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:16:21 +01:00
maximeandClaude Opus 5.5 98c3308ca7 Detector-learner: one benchmark, scripts moved, critic with realistic suggestions (D-120)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 15:12:24 +01:00
maximeandClaude Opus 5.5 98be20a2bc One detector per family, family keys equal to their names (D-119)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:57:52 +01:00
maximeandClaude Opus 5.5 0666d59e7d Detector agents aligned on all labels of labeling/auto/, without events (D-118)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:43:22 +01:00
maximeandClaude Opus 5.5 8ddf539a9f Detector-learner BREITHORN: process fixes
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:41:00 +01:00
maximeandClaude Opus 5.5 516b538cf7 Labels: the trou family and corr_x/corr_y as the anomaly (D-111); thermal movement is always labeled, the camera's included even when compensated (D-117)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:13:15 +01:00
maximeandClaude Opus 5.5 9df0cde364 Detector-learner NEBRASKA: process fixes
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:07:37 +01:00
maximeandClaude Opus 5.5 97253639fa Detector-learner: the global benchmark on all labeled stations at each step, written to detector-learner/historique.csv; the control reads it
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:01:55 +01:00
maximeandClaude Opus 5.5 dbff2b947a Detector-learner: regression.py is the global benchmark on all the labeler's labels, with its history (detector-learner/historique.csv) to compare sessions
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 14:01:32 +01:00
maximeandClaude Opus 5.5 2a8d3da49d Detector-learner: --ameliorer, the improver after the releveur as in /events-learner (D-071); the improver covers both processes
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 13:47:44 +01:00
maximeandClaude Opus 5.5 0537f4ddeb Detector-learner (D-116): the detectors learn from the labeler's labels alone, without events; critic, upgrader and control, the control guarded by the score on the other labeled stations (regression.py); NEBRASKA and NORDEND labels
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 13:46:36 +01:00
maximeandClaude Opus 5.5 9724ea5c3b Presentations: MICHIGAN and NEBRASKA say the redundancy is lost with two references and the correction no longer compensates "correctement"; MICHIGAN's two targets too close since the data start, jump on 5 Oct; sources compile in place in VS Code (brand/ link, Tinymist export on save)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 12:29:37 +01:00
maximeandClaude Opus 5.5 3211c9ce30 Presentations: their sources live next to each report (support/<dossier>/presentation.typ) and compile with .claude/skills/presentation/compiler.sh
The script copies the source and the folder's images into the AIMsight Typst factory, builds, puts the PDF back as presentation.pdf and reports overflowing slides. run.md: how the user edits and compiles a presentation. Skill /presentation, D-114 and CLAUDE.md updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 12:03:03 +01:00
maximeandClaude Opus 5.5 1a04ae0936 Presentations: BREITHORN slide 8 shows the camera's rotation above the targets' curves (the rotation drops on each of the three nights, then comes back); conclusion slides titled "Conclusions" with the bullet points
Skill /presentation: the rotation when it is the angle that moves; the "Conclusions" title; model deck updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 11:58:56 +01:00
maximeandClaude Opus 5.5 44711d2288 Presentations: camera tilt and pan, captions that read on their own, conclusions as bullet points without any instruction to the client
BREITHORN: tilt and pan of the camera (slide 3); targets labelled with their axis in the parallax figure (corr_x horizontal, corr_y vertical); "reflections", not "headlights", on the vehicle; every caption says what each panel shows. MICHIGAN, NEBRASKA: a new figure (tilt, pan, station temperature, horizontal displacement of four targets) for the camera events; complete captions; same conclusion. The client is given facts, not instructions. Skill /presentation and D-114: the rules (no instruction, self-explanatory captions, name only what is visible, "the curves drift", wide multi-panel figures).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 11:57:31 +01:00
maximeandClaude Opus 5.5 a85d736714 Analyst reads the data like the labeler (D-115): same tools and documentation (labels map, families, measurement chain, temperature from metrics/), in analyser-lot and the /labels-learner instructions; its output (etiquettes.json, journal categories, bounds) unchanged for the detectors benchmark; analyser-lot no longer calls itself the loop's truth (D-112)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 11:44:00 +01:00
maximeandClaude Opus 5.5 05e7494e64 Presentations: facts only, nothing asserted about the structure (user: we do not know whether it moves); "Point de situation" on the title page; the camera's movement is real, its cause (fixing, wall or other) not known
BREITHORN, MICHIGAN, NEBRASKA: "the structure does not move", "not a movement of the structure", "these points do not move", "nothing moved" replaced by the observed fact (the camera moves, the jump occurs when the reference is lost, the vehicle hides these targets, the curves follow the camera); conclusions titled by the main facts; "pique du nez" removed from titles and figures. Skill /presentation: the rules and the model deck updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 11:40:46 +01:00
maximeandClaude Opus 5.5 98261833f6 Support: situation reports for BREITHORN, MICHIGAN and NEBRASKA extended to the whole of October 2026, and their presentations; new /presentation skill (D-114)
Reports: all October labels (data from 03.10), the expert's diagnosis kept; new findings before 20.10 (night without measurements 15-16.10 on all three; BREITHORN references hidden on 04, 07, 10.10; MICHIGAN curves agitated 17-18.10; NEBRASKA camera stepping back 13-17.10); figures redone for the month.

Presentations (presentation.pdf in each report folder), built with the AIMsight Typst factory ~/docfactory-presentation, in the structure set by the user: "Point de situation" title with station and period, a numbered chronological summary, one self-sufficient slide per event with its diagnosis and picture (BREITHORN: the vehicle in front of COLI24), a conclusion (why it is less reliable, what to do), closing slide. A diagnosis, never an instruction ("to ignore"); the camera's movement is real, its cause (fixing, wall or other) not known.

Skill .claude/skills/presentation/: structure, the user's rules, what failed, model deck (BREITHORN) and image-de-la-cause.py (brightened crops around a target).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 11:31:17 +01:00
maximeandClaude Opus 5.5 f0ba5b9fcf Labels and journal: "Faulty target" renamed "Target issue" everywhere (user: the name was not clear), key cible_degradee -> cible_a_probleme; labels, events, detectors, Status page, labeling tool, documentation and history (D-109)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 09:31:18 +01:00
maximeandClaude Opus 5.5 026a5bd432 Support skill: /support <STATION> writes a situation report from the labeler's labels (labeling/auto/), one or two images per finding (curves of the targets, camera image when the cause shows), one section per slide; /support <mail> unchanged (D-113)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 17:45:36 +01:00
maximeandClaude Opus 5.5 bc984a86d3 Events-learner on the labeler's labels: the session reads the lot's labels in labeling/auto/ instead of having the analyst read the lot; the detectors-critic judges the detectors against them, family by family (correspondence of labeling/familles.md); score.py --du --au scores by family and by instant, the session page shows labels and detectors on one row per family; analyst, tool-builder and event-builder leave the session (D-112, point 72)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 16:25:44 +01:00
maximeandClaude Opus 5.5 5adb711ce0 Label families: five families for labels, documented in labeling/familles.md and read by the labeler (Visibility, Obstruction, Thermal displacement, Displacement, Faulty target; turbulence merged into Visibility) (D-109, D-110); the labeler keeps only its labels, its work folder goes to out/ (D-106); third trial on WILDSTRUBEL 01-06.09, 30 labels; the labeling tool shows one-measurement labels
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 15:40:35 +01:00
maximeandClaude Opus 5.5 8266293ed3 Labeler agent: an analyst that only produces labels (etiqueteur, D-106), one category per label (D-107), anomalies only and one Thermal displacement category (D-108); second trial on JUNGFRAU 30.09-05.10, 15 labels; the labeling tool gets a Category box above the text, to add and to edit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 14:44:08 +01:00
maximeandClaude Opus 5.5 fe025612cc Automatic labels, first trial: the analyst labels JUNGFRAU 20-25.10 alone with the labels map (7 labels in labeling/auto/, analysis beside them); the labeling tool shows manual or auto labels from a Labels menu (D-105)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 13:40:29 +01:00
maximeandClaude Opus 5.5 d47d29b04e Labels map: the reviewed expert labels arranged by phenomenon for the analyst (labeling/carte-des-etiquettes.md), built by the /labels-map skill with its check script; first pass on 23 labels, 14 phenomena, reviewed against the labels and the skill strengthened (D-104)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 12:47:15 +01:00
maximeandClaude Opus 5.5 a36648c836 JUNGFRAU labels reviewed with the expert (/review-labels): all 8 kept with their bounds, rewritten for clarity; the skill keeps the expert's own labels and only asks on a real objection
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 12:24:34 +01:00
maximeandClaude Opus 5.5 7ca701e502 Label review as the /review-labels skill, one label at a time with the expert, a qualitative text proposed for each, no review file (D-103); NIEDERHORN reviewed: 5 labels kept, 3 train passages removed
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 11:26:34 +01:00
maximeandClaude Opus 5.5 3a07d45ffe Labels in two folders: manual/ (by hand, checked with the label-reviewer, the tool's) and auto/ (what the analyst will find alone, empty) (D-102)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-02 10:27:32 +01:00